\

=
1
log
t
t
t
p
p
R
where p is the price of the S&P500 index. Figure 1 exhibits the
S&P500 return series (R
t
). Tables 1 and 2 show the economic
variables and technical indicators that were employed for the
prediction task. The choice of these primary predictive inputs is
based on their frequent use in the literature [126].
0.03
0.02
0.01
0
0.01
0.02
0.03
Fig. 1. Return series of the S&P500 Index.
Table 1. List of macroeconomic variables
Variables Description
DBAA Moody's Seasoned Baa Corporate Bond Yield
DTB3 3Month Treasury Bill: Secondary Market Rate
DTB6 6Month Treasury Bill: Secondary Market Rate
DFEDTAR Federal Funds Target Rate
DFF Effective Federal Funds Rate
DEXCAUS Canada / U.S. Foreign Exchange Rate
DEXJPUS Japan / U.S. Foreign Exchange Rate
DEXSZUS Switzerland / U.S. Foreign Exchange Rate
DEXUSEU U.S. / Euro Foreign Exchange Rate
DEXUSUK U.S. / U.K Foreign Exchange Rate
DTWEXB Trade Weighted Exchange Index: Broad
DTWEXM Trade Weighted Exchange Index: Major Currencies
Table 2. List of technical indicators
Indicators Description
MA(t) moving average
X1(t) [X(t)  lowest price(t)] / [(highest price(t)  lowest
price(t)]
X2(t) [X(t)  MA(t)] / MA(t)
X3(t) X(t)  MIN[X(t),,X(t5)]
X4(t) X(t)  MAX[X(t),,X(t5)]
MPM(t) [(highest price(t)lowest price(t))(highest price(t
2)lowest price(t2))]/2
BR(t) volume(t)/[(highest price(t)  lowest price(t))/2]
EMV(t) MPM(t)/BR(t)
MOM(t) closing price (t)  closing price (t6)
ROC(t) closing price (t) / closing price (t6)
DIS(t) [closing price (t) / MA(t6)]*100
STOD(t) [closing price(t)  low price(t,t6)]/ [high price(t, t
6) low price(t,t6)]
Unlike previous studies [926] and in order to improve the
coherence of the inputoutput mapping, Granger tests [31]
were performed to identify causal relationships between
predictive inputs and future stock returns. As a result, only
statistically causal inputs are retained and the number of
dimension in the input space is reduced. For instance, the
relevant inputs were selected by first running Granger causality
tests on the macro economic variables and technical indicators.
Since correlation does not necessarily imply causation, the
Granger test was employed to investigate whether x (the macro
variables or the technical indicators) causes y (the stock returns
in this work). This simple statistical approach is about seeing
how the current y can be explained by past values of x and
then whether adding k lagged values of x can improve the
explanation. In other words, y is said to be Grangercaused by
x if x helps forecast y. The Granger test allows considering only
inputs that have a high statistical causal effect on future stock
returns. The test is based on bivariate regressions of the
form:
t k t k y t y k t k x t x x t
t k t k x t x k t k y t y y t
y y x x x
x x y y y
+ + + + + + + =
+ + + + + + + =
, 1 1 , , 1 1 , 0 ,
, 1 1 , , 1 1 , 0 ,
... ...
... ...
where and represent Gaussian disturbances. Then, F
statistics are computed as the Wald statistics [28] for the joint
hypothesis:
0 ...
2 1
= = = =
k
The Fstatistics allows testing whether the coefficients on the
lagged xs are statistically significant in explaining the
dependant y. In this study, the number of lags, k, was set to 5
and the retained statistical significance is 5%. Tables 3 and 4
provide the obtained results from the Granger causality tests
for macroeconomic variables and technical indicators
respectively.
Table 3. Granger tests for macro variables
Null Hypothesis FStat. Probability
DAAA does not Granger Cause R 0.95143 0.4465
DBAA does not Granger Cause R 1.16856 0.32223
DEXCAUS does not Granger Cause R 1.06268 0.37919
DEXJPUS does not Granger Cause R 1.44641 0.20448
DEXSZUS does not Granger Cause R 2.57318 0.02498
DEXUSEU does not Granger Cause R 1.4195 0.21401
DEXUSUK does not Granger Cause R 1.23676 0.2892
DFEDTAR does not Granger Cause R 1.0858 0.36616
DFF does not Granger Cause R 3.07175 0.00911
DTB3 does not Granger Cause R 0.26302 0.93331
International Journal of Computer Applications (0975 8887)
Volume 29 No.3, September 2011
26
DTB6 does not Granger Cause R 1.21703 0.29847
DTWEXB does not Granger Cause R 2.08736 0.06423
DTWEXM does not Granger Cause R 1.72344 0.12589
Table 4. Granger tests for technical indicators
Null Hypothesis FStatistic Probability
BR does not Granger Cause R 0.24284 0.94345
DIS does not Granger Cause R 1.58049 0.16224
EMV does not Granger Cause R 2.22486 0.0494
MOM does not Granger Cause R 1.00126 0.41542
MPM does not Granger Cause R 1.48907 0.1901
ROC does not Granger Cause R 0.98104 0.42785
STOD does not Granger Cause R 0.68897 0.63181
X1 does not Granger Cause R 0.95859 0.44194
X2 does not Granger Cause R 1.58049 0.16224
X3 does not Granger Cause R 1.04415 0.38987
X4 does not Granger Cause R 2.64313 0.02173
Granger causality tests show strong evidence that
DEXSZUS, DFF, EMV, and X4 cause changes in S&P500
returns. For instance, the null hypothesis is rejected with
probability 0.02498 for DEXSZUS, with probability 0.00911
for DFF, with probability 0.0494 for EMV, and with
probability 0.02173 for X4. Thus, two macroeconomic
variables (DEXSZUS and DFF) and two technical indicators
(EMV and X4) are selected to be fed to the classifiers as
inputs. For example, the PNN and SVM are trained and
tested with these resulting data. The first 80% of the data
(1620 observations) were used for model training while
the last 20% (405 observations) were used for outofsample
forecasting. Cross validation is not considered since the goal
of this study is to model time series and no future
observations are used to predict past ones. The variable to be
predicted is stock market future trends. Then, the output
becomes:
{ } 0 1 ; 0 1 > + < =
t t t
R if R if y
In order to test what type of information (macroeconomic
variables versus technical indicators) provide higher accuracy
rate, and whether the combination of the two type of
information helps improving the prediction accuracy, the
following prediction models (PM) are considered:
Model.1:
) , , , , (
1 1 2 1
=
t t k t t t t
DFF DEXSZUS R R R f y K
Model.2:
) , , , , , (
7 1 2 1
X EMV R R R f y
t k t t t t
= K
Model.3:
) , , , , , , , (
7 1 1 1 2 1
X EMV DFF DEXSZUS R R R f y
t t t k t t t t
= K
where t is time script and k is a lag order which is determined
using the autocorrelation function given by:
( )( ) ( )
1
1
2
1
= + =
(
=
T
t
t
T
k t
k t t k
R R R R R R
where R is the sample mean of R
t
. The appropriate k is
determined following the methodology of [32][33]. Indeed, it
is important to include past returns to predict future market
directions if the return series are autocorrelated. In other
words, history of the returns helps predicting future returns.
Figure 2 shows the autocorrelation function up to 10 lags.
Since is nonzero for k=1 and k=2, it means that the series is
serially correlated. In addition, the values of the auto
correlation function die quickly in the first three lags which is a
sign that the series obeys a loworder autoregressive (AR)
process; for example an AR(2). Therefore, the number of lags
to be included in the prediction models is up to k=2. Finally,
the following prediction models (PM) are implemented and
simulated:
PM.1:
) , , , (
1 1 2 1
=
t t t t t
DFF DEXSZUS R R f y
PM.2:
) , , , (
7 1 2 1
X EMV R R f y
t t t t
=
PM.3:
) , , , , , (
7 1 1 1 2 1
X EMV DFF DEXSZUS R R f y
t t t t t t
=
International Journal of Computer Applications (0975 8887)
Volume 29 No.3, September 2011
27
Fig. 2. Autocorrelation function
k
of S&P500 returns at
different lags k [1,10].
2.2 The classifiers
2.2.1 Probabilistic neural networks
The PNN was proposed by Specht [28]. It is built based upon
the Bayesian method of classification. Indeed, the PNN employs
Bayesian decisionmaking theory based on an estimate of the
probability density of the data. The probabilistic neural network
requires only a single presentation of each pattern. The PNN
employs an exponential activation function rather than the
sigmoid function that is commonly used in the multilayer
perceptron. Then, PNN can identify nonlinear decision
boundaries that approach the Bayes optimal [34]. The basic
network topology consists of four layers. The first layer is the
inputs layer. In the second layer, the probability density function
(PDF) of each group of patterns is directly estimated from the
set of training samples using [35] window approximation
method. The third layer performs the summation of all PDFs.
Finally, the Bayesian decision is made in the fourth layer. In
sum, the network structure of PNN is similar to back
propagation neural network; but the main difference is that the
transfer function is replaced by exponential function and all
training samples are stored as weight vectors. For instance, the
PDF is assumed to follows a Gaussian distribution. Then, the
PDF for a feature vector X to be of a certain category A is given
by:
( ) ( ) ( ) ( ) ( )
\

=
m
i
Ai Ai
p p
A
X X X X m X f
1
2 1 5 . 0
2 exp 2 1
where, p is the number of patterns in X, m is the number of the
training patterns of category A, i is the pattern number, and is
the smoothing factor of the Gaussian curves used to construct
the PDF. The value of is optimized during training based on
the clearest separation of classes with the highest classification
rate [36].
2.2.2 Support vector machines
Support Vector Machines (SVM) is a supervised statistical
learning technique introduced by Vapnik [29]. It is one of the
standard tools for machine learning successfully applied in many
different realworld problems. For instance, they have been
successfully applied in financial time series trend prediction
[14][37]. The SVM were originally formulated for binary
classification. The SVM seek to implement an optimal marginal
classifier that minimizes the structural risk in two steps. First,
SVM transform the input to a higher dimensional space with a
kernel (mapping) function. Second, SVM linearly combine them
with a weight vector to obtain the output. As result, SVM
provide very interesting advantages. They avoid local minima in
the optimization process. In addition, they offer scalability and
generalization capabilities. For instance, to solve a binary
classification problem in which the output y{1,+1} SVM seek
for a hyperplane w.(x)+b = 0 to separate the data from classes
+1 and 1 with a maximal margin. Here, x denotes the input
feature vector, w is a weight vector, is the mapping function
to a higher dimension, and b is the bias used for classification of
samples. The maximization of the margin is equivalent to
minimizing the norm of w [38]. Thus, to find w and b, the
following optimization problem is solved:
=
+
n
i
i
C w Minimize
1
2
:
( ) ( ) n i b x w y t s
i i i i
, , 1 0 1 . K = +
where C is a strictly positive parameter that determines the
tradeoff between the maximum margin and the minimum
classification error, n is the total number of samples, and
generalization and is the error magnitude of the classification.
The conditions ensure that no training example should be within
the margins. The number of training errors and examples within
the margins is controlled by the minimization of the term:
=
n
i
i
1
The solution to the previous minimization problem gives the
decision frontier:
( ) ( ) ( ) b x x y x f
i
x
i i i
+ =
where each
i
is a Lagrange coefficient. As mentioned before,
the role of the kernel function is to implicitly map the input
vector into a highdimensional feature space to achieve better
separability. In this study the polynomial kernel is used since it
is a global kernel. For instance, global kernels allow data points
International Journal of Computer Applications (0975 8887)
Volume 29 No.3, September 2011
28
that are far away from each other to have an influence on the
kernel values as well [39].
( ) ( ) ( ) ( ) ( )
d
i i i
x x x x x x K 1 , + = =
where the kernel parameter d is the degree of the polynomial to
be used. In this study, d is set to 2. Finally, the optimal decision
separating function can be obtained as follows:
( )


\

+ =
=
b x x K y sign y
n
i
i i i
1
,
3. RESULTS
The accuracy is calculated based on the number of correct
classifications (Hit Ratio). For example, 100% accuracy is
where all records are properly classified and 0% accuracy is
where none are properly classified. The results obtained from
simulation are shown in Figure 3. They show that the PNN
achieved its highest and lowest accuracy with technical
indicators (54%) and macroeconomic information (53%)
respectively. In addition, the simulations show that the best
performance (64%) is obtained with support vector machines
when macroeconomic information is fed to the classifier. On
the other hand, the lowest performance (49%) is achieved by
support vector machines when technical indicators are used as
inputs to the classifier. The findings suggest that the
performance of each classifier depends on the type of input.
For instance, PNN performs best with technical indicators,
whilst SVM perform best with economic information.
Fig. 3. Classification accuracy
Finally, the performance of both PNN and SVM decreases
when macroeconomic information and technical indicators are
fed to the classifiers. Then, the combination of macroeconomic
information and technical indicators does not help to improve
the prediction accuracy. Although the prediction accuracy is
low, the overall results are interesting since some previous
studies reported that stock prices are approximately close to the
random walk process and; consequently; an accuracy of 56%
in the predictions is a satisfying result for stock forecasting
[40][43]. However, it is possible to improve the accuracy if
we consider two investment strategies. The first strategy is to
predict stock market ups by more than a predetermined rate,
and the second strategy is to detect downs by less than another
predetermined rate. Indeed, investors act as if they extrapolate
a positive price trend by overbuying winners [44]. Therefore,
they are likely to follow a momentum investment strategy and
buy winners [45][47]. Then, in order to improve the prediction
of the future trend in S&P500, two trading strategies are
defined. For instance, the output of PNN is defined according
to two strategies as follows:
Strategy.1:
{ } % 5 . 0
,
1 ; % 5 . 0
,
0
,
< =
t i
R if
t i
R if
t i
y
Strategy.2:
{ } % 5 . 0
,
; % 5 . 0
, ,
0 1 > =
t i
R if
t i
R if
t i
y
The first strategy is designed to detect S&P500 ups by more
than 0.5%. This strategy is designed for an aggressive investor
who is a risktaker. On the other hand, the second strategy is
designed to detect S&P500 downs by less than 0.5%. It is
designed for a riskaverse investor who dislikes losses. The
thresholds 0.5% and 0.5% are arbitrarily chosen. For the
classification task, the PNN is considered since it employs
Bayesian decisionmaking theory which is more suitable to
estimate the probability of future gains (strategy 1) and losses
(strategy 2). The forecasting performance of the PNN given each
strategy is shown in Figure 4.
Fig. 4. Classification accuracy given strategies 1&2
The simulation results show that the accuracy increases in
comparison with the results provided in Figure 3. In addition,
the best accuracy is obtained with macroeconomic information
(88.84%) when strategy 1 is adopted. On the other hand, the
lowest accuracy is obtained with technical indicators (82.79%)
when strategy 2 is adopted. As a result, macroeconomic
information is more suitable for both risktaker and riskaverse
investors. In other words, economic information has proven its
superiority in stock market prediction than technical indicators.
Finally, in both strategies the combination of macroeconomic
information and technical indicators does not help improving
the performance of the prediction system.
International Journal of Computer Applications (0975 8887)
Volume 29 No.3, September 2011
29
4. CONCLUSION
The purpose of this study is to predict stock market trends (ups
and downs) with macroeconomic variables and technical
indicators separately and jointly using intelligent and statistical
systems. Indeed, the contribution is to examine the effect of each
type of predictive inputs; and also their combination; on the
classification accuracy. Another contribution in this study is to
consider a better selection of the predictive variables. For
instance, unlike previous literature [14][48][52] that focus on
the prediction of ups and downs of stock markets; Granger
causality tests are performed to identify inputs (macroeconomic
information and technical indicators) that statistically cause
changes in stock returns. In addition, lagged returns that will
form the input space along with inputs selected by Granger tests
are statistically identified by use of autocorrelation function. In
other words, the number of lagged returns as predictive variables
is not set arbitrarily.
In this work the S&P500 returns are studied through
probabilistic neural networks (PNN) and support vector
machines (SVM). As shown in the experiments, PNN performs
best with technical indicators, whilst SVM perform best with
economic information. In addition, the combination of
macroeconomic information and technical indicators does not
help to improve the prediction accuracy. However, the best
accuracy is obtained accuracy is obtained with SVM using
economic information (64%). Although the prediction accuracy
is low, the overall results are interesting since some previous
studies reported that stock prices follow a random walk process.
To improve the accuracy, two trading strategies are defined. The
first trading strategy is designed for risktaker investors to detect
S&P500 ups. The second trading strategy is designed for risk
averse investors to detect S&P500 downs. As a result, the
accuracy increased up to 88.84% and 83.26% for strategy 1 and
strategy 2 respectively. In addition, the simulations show that for
both strategies, economic information provides more accurate
predictions than technical indicators. Furthermore, the
combination of macroeconomic information and technical
indicators does not help improving the performance of the
prediction system in both trading strategies.
In sum, both PNN and SVM confirm that macroeconomic
information is suitable to predict stock market trends than the use
of technical indicators. This finding is explained by the fact that
economic theory provides strong economic models relating stock
market behaviour to the economy cycles and conditions [18].
On the other hand, technical analysis suffers from lack of
theoretical foundations. For future researches, a large set of stock
markets would be considered to validate the results. Also, the
performance of the SVM with different polynomial orders would
be investigated for each trading strategy.
5. REFERENCES
[1] Castanias R.P. Macro information and the Variability of
Stock Market Prices. Journal of Finance 34 (1979), pp.439
50.
[2] Schwert G William. The Adjustment of Stock Prices to
Information about Inflation. The Journal of Finance 36
(1981), pp.1529.
[3] Schwert G William. Stock Returns and Real Activity: A
Century of Evidence. Journal of Finance 14 (1990),
pp.12371257.
[4] Fama EF. Stock Returns, Real Activity, Inflation and
Money. American Economic Review 71 (1981), pp.71:545
65.
[5] NaiFu Chen, Roll R, Ross R. Economic Forces and The
Stock Market. Journal of Business 59 (1986), pp.383403.
[6] Hardouvelis Gikas A. Macroeconomic Information and
Stock Prices. Journal of Economics and Business
1987;39:131140.
[7] Darrat AF. Stock Returns, Money and Fiscal Deficits.
Journal of Financial and Quantitative Analysis 25 (1990),
pp.38798.
[8] Flannery MJ, Protopapadakis AA. Macroeconomic Factors
Do Influence Aggregate Stock Returns. The R e v i e w o f
F i n a n c i a l Studies 15 (2002), pp.75182.
[9] Yiwen Y, Guizhong L, Zongping Z. Stock market trend
prediction based on neural networks. Multiresolution
Analysis and Dynamical Reconstruction. In Proceedings of
the IEEE/IAFE/INFORMS Conference on Computational
Intelligence for Financial Engineering, 2000.
[10] Wu X, Fung M, Flitman A. Forecasting stock market
performance using hybrid intelligent system. Proceedings
of the International Conference on Computational Science,
2001.
[11] Thawornwong S, Enke D. The adaptive selection of
financial and economic variables for use with artificial
neural networks. Neurocomputing 56 (2004), pp.205232.
[12] Egeli B, Ozturan, M, Badur B. Stock market prediction
using artificial neural networks. Proceedings of 3rd Hawaii
International Conference on Business, 2003.
[13] Chen A S, Leung MT, Daouk H. Application of neural
networks to an emerging financial market: Forecasting and
trading the Taiwan Stock Index. Computers and Operations
Research 30 (2003), pp.901923.
[14] Huang W, Nakamori Y, Wang SY. Forecasting stock
market movement direction with support vector machine.
Computer and Operations Research 32 (2005), pp.2513
2522.
[15] Armano G, Marchesi M, Murru A. A hybrid geneticneural
architecture for stock indexes forecasting. Information
Sciences 17 (2004), pp.333.
[16] Jaruszewicz M, Mandziuk J. One day prediction of
NIKKEI Index considering information from other stock
markets. Lecture Notes in Computer Science 3070 Springer
(2004).
[17] Leigh W, Paz M, Purvis R. An analysis of a hybrid neural
network and pattern recognition technique for predicting
shortterm increases in the NYSE Composite Index. Omega
30 (2002), pp.6976.
[18] Lendasse A, De Bodt E, Wertz V, Verleysen M. Nonlinear
financial time series forecasting Application to the
Belgium 20 Stock Market Index. European Journal of
Economical and Social Systems 14 (2000), pp.8191.
International Journal of Computer Applications (0975 8887)
Volume 29 No.3, September 2011
30
[19] Motiwalla L, Wahab M. Predictable variation and profitable
trading of US equities: A trading simulation using neural
networks. Computer and Operations Research 27 (2000),
pp.11111129.
[20] Bautista C C. Predicting the Philippine Stock Price Index
using artificial neural networks. UPCBA Discussion 0107
(2001).
[21] Dong I, Duan C, Jang MJ. Predicting extreme stock
performance more accurately. A paper written for
Government 2001, 2003.
[22] Atsalakis G, Valavanis K. Neurofuzzy and technical
analysis for stock prediction. Working paper, Technical
University of Crete, 2006.
[23] Baek J, Cho S. Time to jump. Long rising pattern detection
in KOSPI 200 future using an autoassociative neural
network. Lecture Notes in Computer Science 2412.
Springer, 2002.
[24] Dong M, Zhou X. Exploring the fuzzy nature of technical
patterns of U.S. stock market. Proceedings of Fuzzy
System and Knowledge Discovery 1 (2002), pp.324328.
[25] Dourra H, Siy, P. Investment using technical analysis and
fuzzy logic. Fuzzy Sets and Systems, 127 (2002), pp.221
240.
[26] Lam S S. A genetic fuzzy expert system for stock market
timing. Proceedings of the IEEE Conference on
Evolutionary Computation, 2001, pp. 410417.
[27] P. Wasserman. Advanced Methods in Neural Computing.
New York: Van No strand Reinhold, 1993.
[28] D. Specht. Probabilistic Neural Networks. Neural Networks
3 (1990), 109118.
[29] Vapnik V N. The Nature of Statistical Learning Theory,
SpringerVerlag, 1995.
[30] http://research.stlouisfed.org/fred2/
[31] Granger C W J. 1969. Investigating Causal Relations by
Econometric Models and CrossSpectral Methods.
Econometrica 37 (1969), pp.424438.
[32] William H Greene. Econometric Analysis. Prentice Hall;
5th edition, 2002.
[33] Peter J Brockwell, Richard A Davis. Introduction to Time
Series and Forecasting. Springer; 8th Printing. Edition,
(2002).
[34] E. Parzen. On Estimation of a Probability Density Function
and Mode. Annals of Mathematical Statistics 33 (1962),
10651076.
[35] Z.R. Yang, M.B. Platt and H.D Platt. Probabilistic Neural
Networks in Bankruptcy Prediction. Journal of Business
Research 44 (1999), 6774.
[36] C. Gaganis and F. Pasiouras, M. Doumpos. Probabilistic
neural networks for the identification of qualified audit
opinions. Expert Systems with Applications 32 (2007),
114124.
[37] Cao, L. Support vector machines experts for time series
forecasting. Neurocomputing 51 (2003), pp.321339.
[38] N. Cristianini, J.S. Taylor, An Introduction to Support
Vector Machines, Cambridge University Press, Cambridge,
MA, 2000.
[39] Jing Sun, DeSheng Wen, GuangRui Li. An efficient guide
stars classification algorithm via support vector machines.
Second International Conference on Intelligent
Computation Technology and Automation, 2009.
[40] Qian Bo, Khaled Rasheed. Stock market prediction with
multiple classifiers. Applied Intelligence 26 (2007), pp.25
33.
[41] Walczak S. An empirical analysis of data requirements for
financial forecasting with neural networks. Journal of
Management Information Systems 17 (2001), pp.203222.
[42] Tsibouris G, Zeidenberg M. Testing the efficient markets
hypothesis with gradient descent algorithms. In: Refenes
AP (ed) Neural networks in the capital markets. JohnWiley
& Sons, Chichester, England, 1995.
[43] Baestaens D J E, Van Den Bergh W M, Vaudrey H. Market
inefficiencies, technical trading and neural networks. In:
Dunis C; (ed) forecasting financial markets, financial
economics and quantitative analysis. John Wiley & Sons,
Chichester, England, 1996, pp.254260.
[44] T. Odean. Do investors trade too much? American
Economic Review 89 (1999) 103124.
[45] P.A. Andreasson. Explaining the pricevolume relationship:
The difference between price changes and changing prices.
Organizational Behavior and Human Decision Processes 41
(1988), 371389.
[46] W.F.M. DeBondt. Betting on trends: Intuitive forecasts of
financial risk and return. International Journal of
Forecasting 9 (1993), 355371.
[47] H. Hong and J.C. Stein. A unified theory of underreaction,
momentum trading, and overreaction in asset markets.
Journal of Finance 54 (1999), 21432184.
[48] Kim, K. J. Financial time series forecasting using support
vector machines. Neurocomputing 55 (2003), pp.307319.
[49] Tay F E H, Cao L J. Application of support vector
machines in financial time series forecasting. Omega 9
(2001), pp.309317.
[50] Tay F E H, Cao L J. Improved financial time series
forecasting by combining support vector machines with
selforganizing feature map. Intelligent Data Analysis 5
(2001), pp.339354.
[51] Sneha Soni, Shailendra Shrivastava. Classification of
Indian Stock Market Data Using Machine Learning
Algorithms. International Journal on Computer Science and
Engineering 9 (2010), pp.29422946.
[52] Phichhang Ou, Hengshan Wang. Prediction of Stock
Market Index Movement by Ten Data Mining Techniques.
Modern Applied Science 3 (2009), pp.